tao-train-deformable-detr

Pass

Audited by Gen Agent Trust Hub on Jun 30, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill documentation and reference files (e.g., references/tao-deploy-deformable-detr.md) instruct the user to run Docker containers using official NVIDIA images from nvcr.io/nvidia/tao/tao-toolkit. These are well-known and trusted resources provided by the vendor for the skill's intended environment.
  • [PROMPT_INJECTION]: The skill is designed to ingest external data, specifically COCO-formatted JSON annotations and images, which introduces a surface for indirect prompt injection. Malicious instructions could theoretically be embedded in the metadata of these files.
  • Ingestion points: Action definitions in references/skill_info.yaml specify json_file and image_dir as inputs for training, evaluation, and inference tasks.
  • Boundary markers: There are no specific instructions or delimiters mentioned to prevent the agent from potentially following instructions hidden within the ingested dataset files.
  • Capability inventory: The skill invokes the deformable_detr CLI tool to perform various model-related actions.
  • Sanitization: The skill relies on the standard data parsing and validation mechanisms provided by the TAO Toolkit.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 30, 2026, 05:48 PM
Security Audit — agent-trust-hub — tao-train-deformable-detr